A data compression method and system based on machine learning

Abstract

This invention belongs to the field of data compression technology and discloses a data compression method and system based on machine learning. The method includes the following steps: using machine learning algorithms to construct a multimodal feature extraction model and a multimodal data compression model based on several historical multimodal data sets; collecting real-time multimodal data and using the multimodal feature extraction model to extract real-time multimodal data features; updating the multimodal data compression model using a continuous learning algorithm based on the real-time multimodal data features to obtain an updated multimodal data compression model; and compressing the data using the updated multimodal data compression model based on the real-time multimodal data features to obtain real-time compressed data. This invention solves the problems of insufficient generalization ability, poor adaptability to new data, catastrophic forgetting, and high dependence on training data in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of data compression technology, specifically relating to a data compression method and system based on machine learning. Background Technology

[0002] Data compression refers to techniques that reduce the space required for data storage or transmission. With the widespread adoption of the internet, the Internet of Things (IoT), mobile devices, and various sensors, the speed and scale of data generation are exploding. A contradiction exists between limited storage space and the ever-increasing data volume, necessitating compression to save storage resources. Furthermore, network bandwidth limitations require reducing data volume during data transmission to improve speed and efficiency. Data compression can reduce the purchase and maintenance costs of storage hardware, and compressed data can be transmitted faster, improving the real-time performance of data processing.

[0003] Existing data compression technologies have the following drawbacks:

[0004] 1) Insufficient generalization ability: Existing technologies often perform well on specific types of data, but their generalization ability is insufficient for multimodal data. This means that for different types of data, it may be necessary to redesign the compression algorithm, resulting in wasted resources and increased time costs.

[0005] 2) Poor adaptability to new data: Existing data compression techniques are usually based on static models. Once trained, the model is fixed. When new data appears, these models cannot adapt to change, resulting in reduced compression efficiency.

[0006] 3) Catastrophic forgetting exists: Existing data compression models often exhibit catastrophic forgetting when continuously learning new tasks, that is, forgetting previously learned experience during the process of learning new tasks. This is because traditional training methods are usually carried out on fixed datasets and do not take into account the long-term retention of experience.

[0007] 4) High dependence on training data: Existing data compression models usually require a large amount of training data to achieve good performance. When the amount of data is limited, the performance of the model may be greatly reduced. Summary of the Invention

[0008] To address the problems of insufficient generalization ability, poor adaptability to new data, catastrophic forgetting, and high dependence on training data in existing technologies, the present invention aims to provide a data compression method and system based on machine learning.

[0009] The technical solution adopted in this invention is as follows:

[0010] A machine learning-based data compression method includes the following steps:

[0011] Based on several historical multimodal data sets, machine learning algorithms are used to construct a multimodal feature extraction model and a multimodal data compression model.

[0012] Collect real-time multimodal data and use a multimodal feature extraction model to extract real-time multimodal data features.

[0013] Based on the characteristics of real-time multimodal data, a continuous learning algorithm is used to update the multimodal data compression model, resulting in an updated multimodal data compression model.

[0014] Based on the characteristics of real-time multimodal data, an updated multimodal data compression model is used to compress the data and obtain real-time compressed data.

[0015] Furthermore, based on several historical multimodal data sets, machine learning algorithms are used to construct a multimodal feature extraction model and a multimodal data compression model, including the following steps:

[0016] Collect a number of historical multimodal data and preprocess the historical multimodal data to obtain a number of preprocessed historical multimodal data.

[0017] Based on several preprocessed historical multimodal data, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated.

[0018] Using a preset data type as the clustering object, several historical multimodal data features are clustered to obtain several first cluster centers and corresponding historical multimodal data feature clusters;

[0019] Based on the feature clusters of historical multimodal data of several preset data types, an optimized multimodal data compression model is constructed using an adversarial training deep learning algorithm, and several historical data compression experiences are generated.

[0020] Using the experience replay mechanism of continuous learning algorithms, an experience replay pool is set up in the optimized multimodal data compression model, and several historical data compression experiences are stored in the experience replay pool.

[0021] The elastic weight connection mechanism of the continuous learning algorithm is used to adjust the initial loss function of the optimized multimodal data compression model, obtain the adjusted loss function, and obtain the final multimodal data compression model.

[0022] Furthermore, the multimodal feature extraction model is constructed based on the CNN-GCN-LSTM algorithm;

[0023] The multimodal data compression model is constructed based on the cGAN-MLP-DBN algorithm.

[0024] Furthermore, based on several preprocessed historical multimodal data sets, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated, including the following steps:

[0025] The CNN-GCN-LSTM algorithm is used to construct an initial multimodal feature extraction model. The initial multimodal feature extraction model includes a first input layer, an initial image data feature extraction module, an initial graph structure feature extraction module, an initial sequence feature extraction module, and a first output layer.

[0026] The first input layer is used to receive several preprocessed historical multimodal data; each preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data, and preprocessed historical sequence data.

[0027] Based on several preprocessed historical image data, the initial image data feature extraction module is trained to obtain an optimized image data feature extraction module, and several historical image data features are generated.

[0028] Based on several preprocessed historical graph structure data, the initial graph structure data feature extraction module is trained to obtain an optimized graph structure data feature extraction module, and several historical graph structure data features are generated.

[0029] Based on several preprocessed historical sequence data, the initial sequence data feature extraction module is trained to obtain an optimized sequence data feature extraction module, and several historical sequence data features are generated.

[0030] Using the first output layer, the historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data are associated to obtain several historical multimodal data features;

[0031] By integrating the first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module, and the first output layer, the final multimodal feature extraction model is obtained.

[0032] Furthermore, using a preset data type as the clustering object, the FCM clustering algorithm is used to cluster several historical multimodal data features, resulting in several first cluster centers and corresponding historical multimodal data feature clusters.

[0033] Furthermore, based on historical multimodal data feature clusters of several preset data types, an optimized multimodal data compression model is constructed using an adversarial training deep learning algorithm, and several historical data compression experiences are generated, including the following steps:

[0034] The cGAN-MLP-DBN algorithm is used to construct an initial multimodal data compression model. The initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module, and a second output layer.

[0035] The historical multimodal data feature clusters of several preset data types are input into the second input layer, and the initial multimodal data compression module is pre-trained based on the historical multimodal data feature clusters of several preset data types to obtain the pre-trained multimodal data compression module.

[0036] Based on the historical multimodal data feature clusters of several preset data types, the initial adversarial training module is subjected to adversarial training to obtain an optimized adversarial training module, and several extended historical multimodal data features of each preset data type are generated.

[0037] Based on the historical multimodal data feature clusters of several preset data types and several extended historical multimodal data features, the pre-trained multimodal data compression module is trained to obtain an optimized multimodal data compression module and generate several historical data compression experiences.

[0038] The integrated second input layer, optimized adversarial training module, optimized multimodal data compression module, and second output layer yield an optimized multimodal data compression model.

[0039] Furthermore, real-time multimodal data is collected, and a multimodal feature extraction model is used to extract real-time multimodal data features, including the following steps:

[0040] Real-time multimodal data is acquired and preprocessed to obtain preprocessed real-time multimodal data. The preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data.

[0041] The first input layer is used to receive preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data.

[0042] The image data feature extraction module is used to extract real-time image data features from the preprocessed real-time image data.

[0043] The graph structure data feature extraction module is used to extract real-time graph structure data features from the preprocessed real-time graph structure data.

[0044] The sequence data feature extraction module is used to extract real-time sequence data features from the preprocessed real-time sequence data.

[0045] Using the first output layer, real-time image data features, real-time graph structure data features, and real-time sequence data features are integrated to obtain real-time multimodal data features.

[0046] Furthermore, based on the characteristics of real-time multimodal data, a continuous learning algorithm is used to update the multimodal data compression model, resulting in an updated multimodal data compression model, including the following steps:

[0047] Obtain the real-time Euclidean distances between the real-time multimodal data features and several first cluster centers. If all real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step; otherwise, use the current multimodal data compression model to compress the data and obtain real-time compressed data.

[0048] The real-time multimodal data features are set as a new data type other than the preset data type, and the real-time multimodal data features are input into the second input layer of the multimodal data compression model;

[0049] Using the adversarial training module, conditional information is embedded into the features of real-time multimodal data to obtain real-time conditional information embedded features;

[0050] Using the adversarial training module, several real-time training data features belonging to the same new data type as the real-time conditional information embedding features are generated.

[0051] Based on several real-time training data features, the adversarial training module is subjected to adversarial training to obtain an updated adversarial training module.

[0052] Based on several real-time training data features, the multimodal data compression module is trained to generate several real-time data compression experiences.

[0053] Based on the characteristics of real-time multimodal data, several historical data compression experiences are randomly selected from the experience playback pool and mixed with several real-time data compression experiences to obtain several mixed data compression experiences.

[0054] Based on several experiences in mixed data compression, a continuous learning algorithm is used to continuously train the multimodal data compression module, and an adjusted loss function is used to generate real-time loss values ​​during the continuous training process.

[0055] If the real-time loss value is less than the loss threshold, an updated multimodal data compression module is obtained; otherwise, the training of the multimodal data compression module continues.

[0056] By integrating the updated adversarial training module and the updated multimodal data compression module, an updated multimodal data compression model is obtained.

[0057] Furthermore, based on the characteristics of real-time multimodal data, an updated multimodal data compression model is used to compress the data to obtain real-time compressed data, including the following steps:

[0058] The real-time multimodal data features are input into the second input layer of the updated multimodal data compression model, and the real-time multimodal data features are divided into several real-time sub-blocks.

[0059] Several real-time sub-blocks are input into the multimodal data compression module. The multimodal data compression module is used to encode each real-time sub-block to obtain several real-time compressed codewords.

[0060] The second output layer is used to merge several real-time compressed codewords to obtain initial real-time compressed data, and then the initial real-time compressed data is post-processed to obtain the final real-time compressed data.

[0061] A machine learning-based data compression system is provided for implementing a data compression method. The system includes a model building unit, a feature extraction unit, a model update unit, and a data compression unit connected in sequence.

[0062] The beneficial effects of this invention are as follows:

[0063] This invention provides a machine learning-based data compression method and system. The multimodal feature extraction and compression model employed can handle various data types, improving the generalization ability to multimodal data and making it suitable for a wider range of data compression scenarios. The multimodal data compression model constructed using artificial intelligence algorithms can achieve a high compression rate while maintaining high data reconstruction quality, effectively reducing data reconstruction distortion. Furthermore, it can continuously learn and dynamically update the model when faced with new data, adapting to changes in the new data and ensuring that compression efficiency remains optimal, thus improving adaptability to new data. The use of efficient multimodal data feature extraction and compression models reduces unnecessary computational steps and lowers overall computational complexity, making it suitable for deployment in resource-constrained environments. Generative adversarial training generates training data of the same type as the new data in real time, ensuring the continuous learning and training effect of the model even when training data is scarce. It can accurately identify and process new data, and the application of continuous learning algorithms transfers experience learned from historical data to new data, helping to improve learning efficiency and reduce the need for training samples on new data, thus reducing dependence on training data.

[0064] Other beneficial effects of the present invention will be further explained in the specific embodiments. Attached Figure Description

[0065] Figure 1 This is a flowchart of the machine learning-based data compression method in this invention.

[0066] Figure 2 This is a structural block diagram of the machine learning-based data compression system in this invention. Detailed Implementation

[0067] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0068] Example 1:

[0069] like Figure 1 As shown, this embodiment provides a data compression method based on machine learning, including the following steps:

[0070] S1: Based on several historical multimodal data sets, use machine learning algorithms to construct a multimodal feature extraction model and a multimodal data compression model, including the following steps:

[0071] S1-1: Collect several historical multimodal data and preprocess the historical multimodal data to obtain several preprocessed historical multimodal data.

[0072] S1-2: Based on several preprocessed historical multimodal data, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated;

[0073] The multimodal feature extraction model is constructed based on the CNN-GCN-LSTM algorithm, and includes a first input layer, an image data feature extraction module constructed based on the CNN algorithm, a graph structure feature extraction module constructed based on the GCN algorithm, a sequence feature extraction module constructed based on the LSTM algorithm, and a first output layer. The image data feature extraction module, the graph structure feature extraction module, and the sequence feature extraction module are connected in parallel between the first input layer and the first output layer.

[0074] Based on several preprocessed historical multimodal data sets, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated, including the following steps:

[0075] S1-2-1: Using the Convolutional Neural Networks (CNN) - Graph Convolutional Network (GCN) - Long Short-Term Memory (LSTM) algorithm, an initial multimodal feature extraction model is constructed. The initial multimodal feature extraction model includes a first input layer, an initial image data feature extraction module, an initial graph structure feature extraction module, an initial sequence feature extraction module, and a first output layer.

[0076] CNN networks extract local features from image data, such as edges, textures, and shapes, through convolutional and pooling layers, thus obtaining image data features. GCN networks perform feature propagation on graph-structured data through convolution-like operations, extracting node features (positional features, image features, etc.) and edge features representing the positional relationships between nodes from several segmented images, and constructing complex, high-dimensional global graph structure features. LSTM networks process sequence data, such as text, speech, or time series, capturing the temporal dependencies in the data to obtain sequence data features. Through the cooperation of these networks, feature extraction tasks for multimodal data are achieved, providing preprocessing for subsequent data compression, improving the efficiency and speed of data compression, and reducing computational load.

[0077] S1-2-2: Receive several preprocessed historical multimodal data using the first input layer; each preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data, and preprocessed historical sequence data;

[0078] S1-2-3: Based on several preprocessed historical image data, the initial image data feature extraction module is trained to obtain an optimized image data feature extraction module and generate several historical image data features;

[0079] S1-2-4: Based on several preprocessed historical graph structure data, train the initial graph structure data feature extraction module to obtain an optimized graph structure data feature extraction module and generate several historical graph structure data features.

[0080] S1-2-5: Based on several preprocessed historical sequence data, train the initial sequence data feature extraction module to obtain an optimized sequence data feature extraction module and generate several historical sequence data features.

[0081] S1-2-6: Using the first output layer, associate the historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data to obtain several historical multimodal data features;

[0082] S1-2-7: Integrate the first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module, and the first output layer to obtain the final multimodal feature extraction model;

[0083] S1-3: Using a preset data type as the clustering object, the Fuzzy C-means (FCM) clustering algorithm is used to cluster several historical multimodal data features to obtain several first cluster centers and corresponding historical multimodal data feature clusters, including the following steps:

[0084] S1-3-1: Using the FCM clustering algorithm, cluster centers are initialized in several historical multimodal data features to generate several initial first cluster centers;

[0085] S1-3-2: Obtain the initial historical Euclidean distance between each historical multimodal data feature and several initial first cluster centers, and generate the initial historical fuzzy membership degree based on the initial historical Euclidean distance;

[0086] The formula is:

[0087]

[0088] In the formula, For the first i Historical multimodal data features and the first j The initial Euclidean distance of the first cluster centers; For the first Historical multimodal data characteristics; For the first j The initial first cluster centers; For data feature indicators; This is the indicator value for the first cluster center;

[0089]

[0090] In the formula, For the first Historical multimodal data characteristics; For data feature indicators; All of these are initial indicators of the first cluster center; The total number of the first cluster centers; For the first i Historical multimodal data features up to the first j , k The initial distance to the first cluster center; For the first i Historical multimodal data features belong to the first The initial historical fuzzy membership degree of the initial cluster centers;

[0091] S1-3-3: Update the initial first cluster centers based on the initial historical fuzzy membership degree to obtain several updated first cluster centers;

[0092] The formula is:

[0093]

[0094] In the formula, For the first j Updated first cluster center; For fuzzy factors; For data feature indicators; The total number of data; This is the indicator value for the first cluster center; For the first Historical multimodal data characteristics; For the first i Historical multimodal data features belong to the first Updated historical fuzzy membership of the first cluster center;

[0095] S1-3-4: Obtain the updated historical fuzzy membership degree of each historical multimodal data feature and several updated first cluster centers;

[0096] S1-3-5: Based on several updated historical fuzzy membership degrees, the Lagrange multiplier method is used to calculate the merging function, and the merging function value and the change value of the merging function are obtained.

[0097] The formula is:

[0098]

[0099] In the formula, For the first t , t The combined function value of the Lagrange multiplier method after -1 iterations; The corresponding change value; For the first i Feature parameters; t This is an indicator of the number of iterations. For the first i Historical multimodal data features belong to the first Updated historical fuzzy membership of the first cluster center; For fuzzy factors; For data indication; The total number of data; This is the indicator value for the first cluster center; The total number of cluster centers; For the first i Historical multimodal data features up to the first j Updated Euclidean distance of the first cluster centers;

[0100] S1-3-6: If the merge function value is greater than the function threshold, or the change value of the merge function is greater than the change value threshold, then continue to update the first cluster center; otherwise, take the updated first cluster center as the final first cluster center and proceed to the next step.

[0101] S1-3-7: Obtain the final historical Euclidean distance between each historical multimodal data feature and several final first cluster centers, and divide the historical multimodal data features into the final first cluster centers according to the final historical Euclidean distance to obtain several historical multimodal data feature clusters; the historical multimodal data feature clusters include several historical multimodal data features belonging to the same preset data type;

[0102] S1-3-8: Set a corresponding preset data type for each final first cluster center, and spread the preset data type of the final first cluster center to the corresponding historical multimodal data feature cluster;

[0103] S1-4: Based on the historical multimodal data feature clusters of several preset data types, use adversarial training deep learning algorithms to construct an optimized multimodal data compression model and generate several historical data compression experiences;

[0104] The multimodal data compression model is constructed based on the Conditional Generative Adversarial Network (cGAN)-Multilayer Perceptron (MLP)-Deep Belief Network (DBN) algorithm. The multimodal data compression model includes a second input layer, an adversarial training module based on the cGAN-MLP algorithm, a multimodal data compression module based on the DBN algorithm, and a second output layer. The adversarial training module includes a conditional information embedder and a conditional information processor, both based on the MLP algorithm, and a generator and a discriminator, both based on the Recurrent Neural Network (RNN) algorithm. The adversarial training module and the multimodal data compression module are connected in parallel between the first input layer and the first output layer. The generator is connected to the second input layer, the discriminator, and the multimodal data compression module, respectively. The discriminator is connected to the conditional information processor.

[0105] The conditional information embedder processes conditional information, including multimodal data features, to obtain conditional information embedding features in sequence format, integrating multimodal conditional information into the generation process. The generator, based on the sequence data and random noise output by the conditional information embedder, generates several training data features of the same new data type as the conditional information embedding features. The conditional information processor processes additional conditional information on the multimodal data features, helping the discriminator more accurately determine the authenticity of the training data features. The discriminator analyzes whether the generated training data features are authentic and conform to the given conditional information. The generator and discriminator compete against each other through an adversarial training process; the generator attempts to generate training data features that can deceive the discriminator, while the discriminator attempts to better distinguish between real and fake training data features. The adversarial training module performs well in generating realistic data, especially when there is insufficient new data. The generator helps generate training data, achieving sufficient training samples even with limited new data, ensuring the diversity of training samples, enabling continuous training of the multimodal data compression module, enhancing its generalization ability, and thus better handling real new data.

[0106] Based on historical multimodal data feature clusters of several preset data types, an optimized multimodal data compression model is constructed using an adversarial training deep learning algorithm, and several historical data compression experiences are generated, including the following steps:

[0107] S1-4-1: Using the cGAN-MLP-DBN algorithm, construct the initial multimodal data compression model; the initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module, and a second output layer;

[0108] S1-4-2: Input the historical multimodal data feature clusters of several preset data types into the second input layer, and pre-train the initial multimodal data compression module based on the historical multimodal data feature clusters of several preset data types to obtain the pre-trained multimodal data compression module;

[0109] S1-4-3: Based on the historical multimodal data feature clusters of several preset data types, perform adversarial training on the initial adversarial training module to obtain an optimized adversarial training module, and generate several extended historical multimodal data features for each preset data type.

[0110] S1-4-4: Based on the historical multimodal data feature clusters of several preset data types and several extended historical multimodal data features, train the pre-trained multimodal data compression module to obtain an optimized multimodal data compression module and generate several historical data compression experiences.

[0111] S1-4-5: The integrated second input layer, optimized adversarial training module, optimized multimodal data compression module, and second output layer yield an optimized multimodal data compression model;

[0112] S1-5: Using the experience replay mechanism of continuous learning algorithm, set up an experience replay pool in the optimized multimodal data compression model, and store several historical data compression experiences in the experience replay pool;

[0113] The experience replay pool is used to store the interaction experience of the multimodal data compression model in the environment. These experiences are usually stored in the form of (s, a, r, s'), where s is the current state, i.e. the current model parameters of the multimodal data compression model, a is the action taken, including various actions to adjust the model parameters, r is the reward obtained, which is used to characterize the degree of influence of the action on the state, and s' is the next state, i.e. the state after the model parameters are adjusted during the training process.

[0114] Since the experiences in the experience replay pool are randomly sampled, this helps to break the correlation between consecutive experiences, thereby reducing the variance in the training of multimodal data compression models. By reusing experiences, multimodal data compression models can learn more from limited experiences, which is especially useful in the case of new data types or scarce samples. Experience replay helps to stabilize the learning process and reduce fluctuations during training. The size of the experience replay pool is checked periodically, and if it exceeds the preset capacity, some experiences are removed according to a strategy (such as priority sampling, least recently accessed).

[0115] S1-6: Using the elastic weight connection mechanism of the continuous learning algorithm, adjust the initial loss function of the optimized multimodal data compression model to obtain the adjusted loss function and the final multimodal data compression model, including the following steps:

[0116] S1-6-1: Obtain the initial loss function of the optimized multimodal data compression model, as well as several key weights for adjusting the model parameters;

[0117] By calculating the gradient of the weights for task loss, we can identify key weights that are crucial for adjusting model parameters during training. This information can be used to reduce catastrophic forgetting when training new tasks, which is especially important in continuous learning scenarios.

[0118] S1-6-2: Obtain the importance measure of several key weights, and generate a penalty term for adjusting model parameters based on the importance measure;

[0119] To prevent the multimodal data compression model from forgetting previously learned experience when training on a new task, an additional penalty term is added to the loss function using the Elastic Weight Consolidation (EWC) mechanism. This penalty term is proportional to the importance metric of the weights and inversely proportional to the amount of change in the weights on the new data. This penalty term ensures that when training on new data, the weights that are important to the old data do not change too much, thereby reducing the risk of catastrophic forgetting. The larger the importance metric, the more restricted the change of the corresponding weights on the new data. In this way, the multimodal data compression model can retain the experience of the old data while learning new data.

[0120] S1-6-3: Add the penalty term to the initial loss function of the optimized multimodal data compression model to obtain the adjusted loss function, and obtain the final multimodal data compression model;

[0121] S2: Collect real-time multimodal data and use a multimodal feature extraction model to extract real-time multimodal data features, including the following steps:

[0122] S2-1: Acquire real-time multimodal data and preprocess the real-time multimodal data to obtain preprocessed real-time multimodal data; the preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data;

[0123] S2-2: Use the first input layer to receive preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data;

[0124] S2-3: Use the image data feature extraction module to extract real-time image data features from the preprocessed real-time image data;

[0125] S2-4: Use the graph structure data feature extraction module to extract the real-time graph structure data features of the preprocessed real-time graph structure data;

[0126] S2-5: Use the sequence data feature extraction module to extract real-time sequence data features from the preprocessed real-time sequence data;

[0127] S2-6: Using the first output layer, integrate real-time image data features, real-time graph structure data features, and real-time sequence data features to obtain real-time multimodal data features;

[0128] S3: Based on the characteristics of real-time multimodal data, use a continuous learning algorithm to update the multimodal data compression model, obtaining an updated multimodal data compression model, including the following steps:

[0129] S3-1: Obtain the real-time Euclidean distance between the real-time multimodal data features and several first cluster centers. If all real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step. Otherwise, use the current multimodal data compression model to compress the multimodal data compression model and obtain real-time compressed data.

[0130] To automatically identify the data characteristics of new data types, a continuous training environment for the multimodal data compression model is prepared in advance, including the preparation of training samples, so that the multimodal data compression model can accurately process the new data characteristics of new data types.

[0131] S3-2: Set the real-time multimodal data features to a new data type other than the preset data type, and input the real-time multimodal data features into the second input layer of the multimodal data compression model;

[0132] S3-3: Using the conditional information embedder of the adversarial training module, conditional information is embedded into the real-time multimodal data features to obtain real-time conditional information embedded features;

[0133] S3-4: Using the generator of the adversarial training module, generate several real-time training data features that belong to the same new data type as the real-time conditional information embedding features.

[0134] S3-5: Based on several real-time training data features, perform adversarial training on the adversarial training module to obtain an updated adversarial training module;

[0135] S3-6: Based on several real-time training data features, train the multimodal data compression module to generate several real-time data compression experiences;

[0136] S3-7: Based on the characteristics of real-time multimodal data, randomly select several historical data compression experiences from the experience playback pool and mix them with several real-time data compression experiences to obtain several mixed data compression experiences.

[0137] By incorporating historical data compression experience, multimodal data compression models can transfer lessons learned from previous data to new data. This helps improve learning efficiency and reduce the training sample size required for new data. Combining old experience can help prevent multimodal data compression models from forgetting the lessons learned on old data when training on new data. This is a common challenge in continuous learning, known as "catastrophic forgetting." Old experience can provide a broader view of the data distribution, which helps multimodal data compression models generalize better on new data.

[0138] S3-8: Based on several experiences in mixed data compression, a continuous learning algorithm is used to continuously train the multimodal data compression module, and an adjusted loss function is used to generate real-time loss values ​​during the continuous training process.

[0139] S3-9: If the real-time loss value is less than the loss value threshold, the updated multimodal data compression module is obtained; otherwise, the training of the multimodal data compression module continues.

[0140] S3-10: Integrate the updated adversarial training module and the updated multimodal data compression module to obtain an updated multimodal data compression model;

[0141] S4: Based on the characteristics of real-time multimodal data, use an updated multimodal data compression model to compress the data and obtain real-time compressed data, including the following steps:

[0142] S4-1: Input the real-time multimodal data features into the second input layer of the updated multimodal data compression model, and divide the real-time multimodal data features into several real-time sub-blocks;

[0143] S4-2: Input several real-time sub-blocks into the multimodal data compression module, and use the multimodal data compression module to encode each real-time sub-block to obtain several real-time compressed codewords;

[0144] The multimodal data compression module is equipped with multiple compression channels, which can process data features of different modalities simultaneously, improving its adaptability to multimodal data scenarios.

[0145] S4-3: Using the second output layer, merge several real-time compressed codewords to obtain the initial real-time compressed data, and then post-process the initial real-time compressed data to obtain the final real-time compressed data.

[0146] Through pre-training, DBN networks have learned effective representations of multimodal data features and can compress data based on these features, thereby enabling compressed data to retain information while occupying less storage space or bandwidth.

[0147] Example 2:

[0148] like Figure 2 As shown, this embodiment provides a machine learning-based data compression system for implementing a data compression method. The system includes a model building unit, a feature extraction unit, a model update unit, and a data compression unit connected in sequence.

[0149] The model building unit is used to construct a multimodal feature extraction model and a multimodal data compression model based on several historical multimodal data sets and using machine learning algorithms.

[0150] The feature extraction unit is used to collect real-time multimodal data and use a multimodal feature extraction model to extract real-time multimodal data features.

[0151] The model update unit is used to update the multimodal data compression model based on the characteristics of real-time multimodal data and using a continuous learning algorithm to obtain the updated multimodal data compression model.

[0152] The data compression unit is used to compress data based on the characteristics of real-time multimodal data and using an updated multimodal data compression model to obtain real-time compressed data.

[0153] This invention provides a machine learning-based data compression method and system. The multimodal feature extraction and compression model employed can handle various data types, improving the generalization ability to multimodal data and making it suitable for a wider range of data compression scenarios. The multimodal data compression model constructed using artificial intelligence algorithms can achieve a high compression rate while maintaining high data reconstruction quality, effectively reducing data reconstruction distortion. Furthermore, it can continuously learn and dynamically update the model when faced with new data, adapting to changes in the new data and ensuring that compression efficiency remains optimal, thus improving adaptability to new data. The use of efficient multimodal data feature extraction and compression models reduces unnecessary computational steps and lowers overall computational complexity, making it suitable for deployment in resource-constrained environments. Generative adversarial training generates training data of the same type as the new data in real time, ensuring the continuous learning and training effect of the model even when training data is scarce. It can accurately identify and process new data, and the application of continuous learning algorithms transfers experience learned from historical data to new data, helping to improve learning efficiency and reduce the need for training samples on new data, thus reducing dependence on training data.

[0154] This invention is not limited to the optional embodiments described above, and anyone can derive other various forms of products based on the inspiration of this invention. The specific embodiments described above should not be construed as limiting the scope of protection of this invention; the scope of protection of this invention should be determined by the claims, and the specification can be used to interpret the claims.

Claims

1. A data compression method based on machine learning, characterized in that: Includes the following steps: Based on historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using machine learning algorithms, including the following steps: Collect a number of historical multimodal data and preprocess the historical multimodal data to obtain a number of preprocessed historical multimodal data. Based on several preprocessed historical multimodal data, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated. The multimodal feature extraction model is constructed based on the CNN-GCN-LSTM algorithm; the multimodal feature extraction model includes a first input layer, an image data feature extraction module, a graph structure feature extraction module, a sequence feature extraction module, and a first output layer; Using a preset data type as the clustering object, several historical multimodal data features are clustered to obtain several first cluster centers and corresponding historical multimodal data feature clusters; Based on the feature clusters of historical multimodal data of several preset data types, an optimized multimodal data compression model is constructed using an adversarial training deep learning algorithm, and several historical data compression experiences are generated. The multimodal data compression model described above is constructed based on the cGAN-MLP-DBN algorithm; Using the experience replay mechanism of continuous learning algorithms, an experience replay pool is set up in the optimized multimodal data compression model, and several historical data compression experiences are stored in the experience replay pool. The elastic weight connection mechanism of the continuous learning algorithm is used to adjust the initial loss function of the optimized multimodal data compression model, obtain the adjusted loss function, and obtain the final multimodal data compression model. Collect real-time multimodal data and use a multimodal feature extraction model to extract real-time multimodal data features. Based on the characteristics of real-time multimodal data, a continuous learning algorithm is used to update the multimodal data compression model, resulting in an updated multimodal data compression model. Based on the characteristics of real-time multimodal data, an updated multimodal data compression model is used to compress the data and obtain real-time compressed data.

2. The data compression method based on machine learning according to claim 1, characterized in that: Based on several preprocessed historical multimodal data sets, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and several historical multimodal data features are generated, including the following steps: An initial multimodal feature extraction model is constructed using the CNN-GCN-LSTM algorithm. The initial multimodal feature extraction model includes a first input layer, an initial image data feature extraction module, an initial graph structure feature extraction module, an initial sequence feature extraction module, and a first output layer. The first input layer is used to receive a plurality of preprocessed historical multimodal data; each of the preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data, and preprocessed historical sequence data; Based on several preprocessed historical image data, the initial image data feature extraction module is trained to obtain an optimized image data feature extraction module, and several historical image data features are generated. Based on several preprocessed historical graph structure data, the initial graph structure data feature extraction module is trained to obtain an optimized graph structure data feature extraction module, and several historical graph structure data features are generated. Based on several preprocessed historical sequence data, the initial sequence data feature extraction module is trained to obtain an optimized sequence data feature extraction module, and several historical sequence data features are generated. Using the first output layer, the historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data are associated to obtain several historical multimodal data features; By integrating the first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module, and the first output layer, the final multimodal feature extraction model is obtained.

3. The data compression method based on machine learning according to claim 2, characterized in that: Using a preset data type as the clustering object, the FCM clustering algorithm is used to cluster several historical multimodal data features, resulting in several first cluster centers and corresponding historical multimodal data feature clusters.

4. The data compression method based on machine learning according to claim 3, characterized in that: Based on historical multimodal data feature clusters of several preset data types, an optimized multimodal data compression model is constructed using an adversarial training deep learning algorithm, and several historical data compression experiences are generated, including the following steps: An initial multimodal data compression model is constructed using the cGAN-MLP-DBN algorithm. The initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module, and a second output layer. The historical multimodal data feature clusters of several preset data types are input into the second input layer, and the initial multimodal data compression module is pre-trained based on the historical multimodal data feature clusters of several preset data types to obtain the pre-trained multimodal data compression module. Based on the historical multimodal data feature clusters of several preset data types, the initial adversarial training module is subjected to adversarial training to obtain an optimized adversarial training module, and several extended historical multimodal data features of each preset data type are generated. Based on the historical multimodal data feature clusters of several preset data types and several extended historical multimodal data features, the pre-trained multimodal data compression module is trained to obtain an optimized multimodal data compression module and generate several historical data compression experiences. The integrated second input layer, optimized adversarial training module, optimized multimodal data compression module, and second output layer yield an optimized multimodal data compression model.

5. The data compression method based on machine learning according to claim 4, characterized in that: Collect real-time multimodal data and use a multimodal feature extraction model to extract real-time multimodal data features, including the following steps: Real-time multimodal data is acquired and preprocessed to obtain preprocessed real-time multimodal data; the preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data. The first input layer is used to receive preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data. The image data feature extraction module is used to extract real-time image data features from the preprocessed real-time image data. The graph structure data feature extraction module is used to extract real-time graph structure data features from the preprocessed real-time graph structure data. The sequence data feature extraction module is used to extract real-time sequence data features from the preprocessed real-time sequence data. Using the first output layer, real-time image data features, real-time graph structure data features, and real-time sequence data features are integrated to obtain real-time multimodal data features.

6. The data compression method based on machine learning according to claim 5, characterized in that: Based on the characteristics of real-time multimodal data, a continuous learning algorithm is used to update the multimodal data compression model, resulting in an updated multimodal data compression model. The steps include: Obtain the real-time Euclidean distances between the real-time multimodal data features and several first cluster centers. If all real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step; otherwise, use the current multimodal data compression model to compress the data and obtain real-time compressed data. The real-time multimodal data features are set as a new data type other than the preset data type, and the real-time multimodal data features are input into the second input layer of the multimodal data compression model; Using the adversarial training module, conditional information is embedded into the features of real-time multimodal data to obtain real-time conditional information embedded features; Using the adversarial training module, several real-time training data features belonging to the same new data type as the real-time conditional information embedding features are generated. Based on several real-time training data features, the adversarial training module is subjected to adversarial training to obtain an updated adversarial training module. Based on several real-time training data features, the multimodal data compression module is trained to generate several real-time data compression experiences. Based on the characteristics of real-time multimodal data, several historical data compression experiences are randomly selected from the experience playback pool and mixed with several real-time data compression experiences to obtain several mixed data compression experiences. Based on several experiences in mixed data compression, a continuous learning algorithm is used to continuously train the multimodal data compression module, and an adjusted loss function is used to generate real-time loss values ​​during the continuous training process. If the real-time loss value is less than the loss value threshold, an updated multimodal data compression module is obtained; otherwise, the training of the multimodal data compression module continues. By integrating the updated adversarial training module and the updated multimodal data compression module, an updated multimodal data compression model is obtained.

7. The data compression method based on machine learning according to claim 6, characterized in that: Based on the characteristics of real-time multimodal data, an updated multimodal data compression model is used to compress the data to obtain real-time compressed data, including the following steps: The real-time multimodal data features are input into the second input layer of the updated multimodal data compression model, and the real-time multimodal data features are divided into several real-time sub-blocks. Several real-time sub-blocks are input into the multimodal data compression module. The multimodal data compression module is used to encode each real-time sub-block to obtain several real-time compressed codewords. The second output layer is used to merge several real-time compressed codewords to obtain initial real-time compressed data, and then the initial real-time compressed data is post-processed to obtain the final real-time compressed data.

8. A machine learning-based data compression system for implementing the data compression method as described in any one of claims 1-7, characterized in that: The system includes a model building unit, a feature extraction unit, a model updating unit, and a data compression unit connected in sequence.

Images